Brief bio
I am interested in probabilistic approaches to deep learning - especially deep generative models of
sequential data and machine learning applied in health technology and climate applications.
Currently, I primarily research how to model extreme weather events by introducing modelling components that draws upon stochastic dynamical systems.
Previously, my research has focused on how we can make audio models that generalize to real-world use scenarios.
I've been working on variational autoencoders and how probabilistic modelling enables us
to, e.g., learn robust latent representations, incorporate prior knowledge, and utilize uncertainty
quantification.
Below is a collection of some recent work and news.
I started as a postdoc at UC Berkeley BAIR and ICSI with A. Krishnapriyan and M. Mahoney. I'll continue my work on neural stochastic differential equations focusing on modelling extreme weather events using heavy-tailed diffusions.
In Januar 2023, I finished my PhD. My thesis is titled: On Learning Useful Variational Autoencoder Representations. My advisors were M. Moerup (DTU), J. B. B. Nielsen (WSA), A. Kressner (DTU), A. Westermann (WSA), and L. K. Hansen (DTU).
In November 2022, I was invited to give a talk on speaker separation at the Danish Sound Cluster, video here and slides here.
I visited Berkeley working with M. Mahoney. We developed continuously deep hierarchical variational autoencoders by combining neural stochstic differential equations with hierarchical variational autoencoder models.
Multi-subject, multi-modal modelling of functional neuroimaging data using directional statistics. Published in Frontiers in Neuroscience (paper, Twitter thread).
The Variational Inference Time-domain Audio Separation Network (VI-TasNet) is probabilistic extension of TasNets. The work shows how we can incorporate domain-knowledge priors in modelling and quantify separation performance uncertainty unintrusively. We also investigate how speaker separation generalization can be understood through the lense of rate-distortion analysis.
(preprint available)
We present a federated learning approach for learning a client adaptable, robust model when data is non-identically and non-independently distributed (non-IID) across clients, as well as a way to simulate non-IID clients.
(arXiv, presented at FL-ICML 2020 )
I organized and helped teach a course on deep unsupervised learning at DTU, modelled on the Berkeley CS294-158.
Together with WSAudiology and DTU CogSys , I received 60k EUR industrial PhD scholarship from the Innovation Fund Denmark.
Teaching high schools students about sound and machine learning using Google Colab (see export on GitHub here ). Photo by Forskningens Døgn.
I developed a series of machine learning demos for an introductory machine learning course at DTU .
The work was done for my Master's thesis, as a visiting student research at Stanford University (School of Medicine) in collaboration between DTU, Stanford Center for Sleep Sciences and Medicine, and Rigshospitalet Denmark .
We received a grant of 7k EUR for the production of a documentary for an alternative view on a developing country. Through Engineering World Health at DTU I was deployed at Okhaldhunga Community Hospitals. My work at the hospital entailed hospital equipment repair, healthcare staff training (proper use and maintenance), and needfinding (initial phase design research and planning). Photo by S. Sundgaaard.